View
4
Download
0
Category
Preview:
Citation preview
A Theory of Credit Scoring and
Competitive Pricing of Default Risk1
Satyajit Chatterjee
Federal Reserve Bank of Philadelphia
Dean Corbae
University of Texas at Austin
Jose-Vıctor Rıos-Rull
University of Pennsylvania, University of Minnesota, and CAERP
August 2009 (First Draft: 2004)
1The authors wish to thank Kartik Athreya, Hal Cole, Igor Livshits, Jim MacGee, and BorghanNarajabad for helpful comments, as well as seminar participants at the Bank of England, Cam-bridge, CEMFI, Iowa, the Federal Reserve Banks of Atlanta, New York and Richmond, NYU, OhioState, UCLA, USC, UC Riverside, Virginia,Western Ontario, the Philadelphia MacroWorkshop,the Econometric Society Summer Meetings, the SED meetings, and the Fifth PIER-IGIER Con-ference. Chatterjee and Corbae also wish to thank the FRB Chicago for hosting them as visitors.Corbae wishes to thank the National Science Foundation for support under SES-0751380. Finally,we wish to thank Pablo D’Erasmo and Daphne Chen for outstanding research assistance.
Abstract
We propose a theory of unsecured consumer credit where: (i) borrowers have the legal option
to default; (ii) defaulters are not exogenously excluded from future borrowing; (iii) there is
free entry of lenders; and (iv) lenders cannot collude to punish defaulters. In our framework,
limited credit or credit at higher interest rates following default arises from the lender’s
optimal response to limited information about the agent’s type and earnings realizations.
The lender learns from an individual’s borrowing and repayment behavior about his type and
encapsulates his reputation for not defaulting in a credit score. We take the theory to data
estimating the parameters of the model to match key data moments such as the overall and
subprime delinquency rates. We test the theory by showing that our underlying framework
is broadly consistent with the way credit scores affect unsecured consumer credit market
behavior. The framework can be used to shed light on household consumption smoothing
with respect to transitory income shocks and to examine the welfare consequences of legal
restrictions on the length of time adverse events can remain on one’s credit record.
1 Introduction
It is well known that lenders use credit scores to regulate the extension of consumer credit.
People with high scores are offered credit on more favorable terms. People who default
on their loans experience a decline in their scores and, therefore, lose access to credit on
favorable terms. People who run up debt also experience a decline in their credit scores and
have to pay higher interest rates on new loans. While credit scores play an important role
in the allocation of consumer credit, credit scoring has not been adequately integrated into
the theoretical literature on consumption smoothing and asset pricing. This paper attempts
to remedy this gap.1
We propose a theory of unsecured consumer credit where: (i) borrowers have the legal
option to default; (ii) defaulters are not exogenously excluded from future borrowing; (iii)
there is free entry of lenders; and (iv) lenders cannot collude to punish defaulters. We
use the framework to try to understand why households typically face limited credit or
credit at higher interest rates following default and why this changes over time. We show
such outcomes arise from the lender’s optimal response to limited information about the
agent’s type and earnings realizations. The lender learns from an individual’s borrowing and
repayment behavior about his type and encapsulates his reputation for not defaulting in a
credit score.
The legal environment surrounding the U.S. unsecured consumer credit market is char-
acterized by the following features. Individual debtors have can file for bankruptcy under
Chapter 7 which permanently discharges net debt (liabilities minus assets above statewide
exemption levels). A Chapter 7 filer is ineligible for a subsequent Chapter 7 discharge for 6
years. During that period, the individual is forced into Chapter 13 which is typically a 3-5
year repayment schedule followed by discharge. Over two-thirds of household bankruptcies
1One important attempt to remedy this deficiency in the consumption smoothing literature is Gross andSouleles ¡cite¿GSQJE¡/cite¿. That paper empirically tests whether consumption is excessively sensitive tovariations in credit limits taking into account a household’s risk characteristics embodied by credit scores.
1
in the U.S. are Chapter 7. The Fair Credit Reporting Act requires credit bureaus to exclude
the filing from credit reports after 10 years (and all other adverse items after 7 years).
Beginning with the work of Athreya [1], there has been a growing number of papers that
have tried to understand bankruptcy data using quantitative, heterogeneous agent models
(for example Chatterjee, et. al. [4], Livshits, et. al. [14]). For simplicity, these models have
assumed that an individual is exogenously excluded from borrowing while a bankrutpcy
remains on his credit record. This exclusion restriction is often modelled as a Markov
process and calibrated so that on average the household is excluded for 10 years, after which
the Fair Credit Reporting Act requires that it be stricken from the household’s record.
This assumption is roughly consistent with the findings by Musto [16] who documents the
following important facts: (1) households with low credit ratings face very limited credit
lines (averaging around $215) prior to and $600 following the removal of a bankruptcy flag;
(2) for households with medium and high credit ratings, their average credit lines were a
little over $800 and $2000 respectively prior to the year their bankruptcy flag was removed
from their record; and (3) for households with high and medium credit ratings, their average
credit lines jumped nearly doubled to $2,810 and $4,578 in the year that the bankruptcy flag
was removed from their record.2
While this exogenous exclusion restriction is broadly consistent with the empirical facts, a
fundamental question remains. Since a Chapter 7 filer is ineligible for a subsequent Chapter 7
discharge for 6 years (and at worst forced into a subsequent Chapter 13 repayment schedule),
why don’t we see more lending to those who declare bankruptcy? If lenders believe that
the Chapter 7 bankruptcy signals something relatively permanent about the household’s
unobservable characteristics, then it may be optimal for lenders to limit future credit. But if
the circumstances surrounding bankruptcy are temporary (like a transitory, adverse income
shock), those individuals who have just shed their previous obligations may be a good future
credit risk. Competitive lenders use current repayment and bankruptcy status to try to infer
2These numbers are actually drawn from Table III, panel A of Musto’s Wharton working paper #99-22.
2
an individual’s future likelihood of default in order to correctly price loans. There is virtually
no existing work embedding this inference problem into a quantitative, dynamic model.
Figure 1: Delinquency Rates in the Population
up to 499 500−549 550−599 600−649 650−699 700−749 749−800 800+0
10
20
30
40
50
60
70
80
90
100
FICO score range
Delinquency rates by FICO score
delinquency rate (%)% people
15%
51%
31%
15%18%
27%
13%
2%5%
8%5%
2% 1%
12%
87%
71%
Given commitment frictions, it’s important for a lender to assess the probability that a
borrower will fail to pay back – that is, assess the risk of default. In the U.S., lenders use
credit scores as an index of the risk of default. The credit scores most commonly used are
produced by a single company, the Fair Isaac and Company, and are known as FICO scores.3
These scores range between 300 and 850, where a higher score signals a lower probability
of default. Scores under 620, which account for roughly one quarter of the population with
scores, are called “subprime”.4 There is ample empirical evidence that households with
subprime credit scores are more likely to default. Figure 1 provides one such example. As
discipline on our theory, we require our framework to match key credit market facts like that
in Figure 1.
3Over 75% of mortgage lenders and 80% of the largest financial institutions use FICO scores in theirevaluation and approvals process for credit applications.
4http://www.privacyrights.org/fs/fs6c-CreditScores.htm.
3
A FICO score aggregates information from an individual’s credit record like his payment
history (most particularly the presence of adverse public records such as bankruptcy and
delinquency) and current amounts owed.5 It’s also worth noting the kinds of information
that are not used in credit scores. By law, credit scores cannot use information on race,
color, national origin, sex, and marital status. Further, FICO scores do not use age, assets,
salary, occupation, and employment history.
These scores appear to affect the extension of consumer credit in four primary ways.
1. Credit terms (e.g. interest rates) improve with a person’s credit score.
2. The presence of adverse public records (e.g. a bankruptcy) lowers an individual’s score
and removal can substantially raise it.
3. Taking on more debt (paying off debt) tends to lower (raise) credit scores.
4. Credit scores are mean reverting.
The Fico website (http://www.myfico.com/myfico/Credit Central/LoanRates.asp) doc-
uments the negative relationship between FICO scores and average interest rates on loans.
Item 2 is consistent with evidence provided in Musto [16], as well as Fisher, et. al. [9].
Using data from the PSID and SCF, Fisher, et. al. document that a higher percentage of
post-bankruptcy households were denied access to credit. Musto found (p.735) “there is a
strong tenth year effect for the best initial credits...these consumers move ahead of 19% of
the nonfiler population in apparent creditworthiness when their flags are removed.” Further-
more, he states (p.740) “...the boost translates to significant new credit access for these filers
over the ensuing year”. Items 1 and 2 taken together imply that an individual who fails to
pay back an unsecured loan will experience an adverse change in the terms of (unsecured)
credit. Thus, a failure to pay back a loan adversely impacts the terms of credit and may
5The score also takes into account the length of a person’s credit history, the kinds of credit accounts(retail credit, installment credit etc.) and the borrowing capacity (or line of credit) on each account.
4
result in outright denial of credit. Item 3 is consistent with the advice given by FICO for
improving one’s credit score.6 Item 3 in conjunction with item 1 indicates that even absent
default, the terms of credit on unsecured credit worsen as an individual gets further into
debt – people face a rising marginal cost of funds. Item 4 is documented by Musto [16].
These facts suggest the following characterization of the workings of the unsecured con-
sumer credit market. Given the inability of borrowers to commit to pay back, lenders condi-
tion the terms of credit (including whether they lend at all) on an individual’s credit history
encapsulated by a credit score. Individuals with higher scores are viewed by lenders as less
likely to default and receive credit on more attractive terms. A default may signal something
about the borrower’s future ability to repay and leads to a drop in the individual’s credit
score. Consequently, post-default access to credit is available on worse terms and may not
be available at all. Even absent default, greater indebtedness may signal something about
the borrower’s future ability to repay which subsequently leads to a lower credit score and
worse terms of credit.
There is now a fairly substantial literature (beginning with Kehoe and Levine [13]) on
how and to what extent borrowing can occur when agents cannot commit to pay back. This
literature typically assumes that a default triggers permanent exclusion from credit mar-
kets. A challenge for this literature is to specify a structure with free entry of lenders and
where lenders cannot collude to punish defaulters that can make quantitative sense of the
characterization of a competitive unsecured consumer credit market with on-the-equilibrium-
path default offered in the previous paragraphs. This paper take steps toward meeting this
challenge.7 We consider an environment with a continuum of infinitely-lived agents who at
any point in time may be one of two types that affect their earnings realizations and prefer-
ences. An agent’s type is drawn independently from others and follows a persistent two-state
6To improve a score, FICO advises to “Keep balances low on credit card and ‘other revolving credit”’and “[p]ay off debt rather than moving it around”. Source:www.myfico.com/CreditEducation/ImproveYourScore
7In Chatterjee, et.al. [5] we show that credit can be supported even in a finite horizon model wheretrigger strategies cannot support credit.
5
Markov process. Importantly, a person’s type and earnings realizations are unobservable to
the lender.8
These people interact with competitive financial intermediaries that can borrow in the
international credit market at some fixed risk-free rate and make one-period loans to indi-
viduals at an interest rate that reflects that person’s risk of default.9 Because differences
in earnings distributions and preferences bear on the willingness of each type of agent to
default, intermediaries must form some assessment of a person’s type which is an input into
his credit score. We model this assessment as a Bayesian inference problem: intermediaries
use the recorded history of a person’s actions in the credit market to update their prior
probability of his or her type and then charge an interest rate that is appropriate for that
posterior. The fundamental inference problem for the lender is to assess whether a borrower
or a defaulter is a chronically “risky” type or just experiencing a temporary shortfall in earn-
ings. A rational expectations equilibrium requires that a lender’s perceived probability of
an agent’s default must equal the objective probability implied by the agent’s decision rule.
Incorporating this equilibrium Bayesian credit scoring function into a dynamic incomplete
markets model is the main technical challenge of our paper.
We model the pricing of unsecured consumer loans in the same fashion as in our pre-
decessor paper Chatterjee, et.al. [4]. As in that paper, all one-period loans are viewed as
discount bonds and the price of these bonds depend on the size of the bond. This is necessary
because the probability of default (for any type) will depend on the size of the bond (i.e.,
on the person’s liability). If the bond price is independent of the size of the loan and other
characteristics, as it is in Athreya [1], then large loans which are more likely to be defaulted
upon must be subsidized by small loans which are less likely to be defaulted upon. But with
competitive credit markets, such cross subsidization of pooling contracts will fail to be an
equilibrium. This reasoning is corroborated by recent empirical work by Edelberg [8] who
8Ausubel [2] documents adverse selection in the credit market both with respect to observable and un-observable household characteristics.
9Our earlier paper Chatterjee, et. al. [4] shows that there is not a big gain to relaxing the fixed risk-freerate assumption.
6
finds that there has been a sharp increase in the cross-sectional variance of interest rates
charged to consumers.
In Chatterjee, et.al. [4], we also assumed that the price of a one-period bond depended
on certain observable household type characteristics like whether households were blue or
white collar workers. Here we assume those characteristics are not observable but instead
assume that the bond depends on the agent’s probability of repayment, in other words, his
credit score. The probability of repayment depends on the posterior probability of a person
being of a given type conditional on selling that particular sized bond. This is necessary
because the two types will not have the same probability of default for any given sized bond
and a person’s asset choice is potentially informative about the person’s type. With this
asset market structure, competition implies that the expected rate of return on each type of
bond is equal to the (exogenous) risk-free rate.
This is possibly the simplest environment one could imagine that could make sense of
the observed connection between credit history and the terms of credit. Suppose it turns
out that, in equilibrium, one type of person, say type g, has a lower probability of default.
Then, under competition, the price of a discount bond (of any size) could be expected to be
positively related to the probability of a person being of type g. Further, default will lower
the posterior probability of being of type g because type g people default less frequently.
This provides the basis for a theory why people with high scores are offered credit on more
favorable terms.This would explain the fact that people with high scores are offered credit
on more favorable terms.
There are two strands of existing literature to which our paper is closely related. The
first strand relates to Diamond’s [7] well-known paper on acquisition of reputation in debt
markets. Besides differences in the environment (e.g. preferences in his case are risk neutral),
the main difference is that here the decision to default is endogenous while in Diamond it
happens exogenously. The second strand relates to the paper of Cole, Dow and English [6]
on sovereign debt. In their setting a sovereign who defaults is shut out of international credit
7
markets until such time as the sovereign makes a payment on the defaulted debt. Chapter
7 bankruptcy law, which we consider here, results in discharge of uncollateralized debt.10
Further, the law does not permit individuals to simultaneously accumulate assets during the
discharge of debt granted by the bankruptcy court.11
Our framework has the ability to address an interesting question that arises from Musto’s
empirical work. What are the effects on consumption smoothing and welfare of imposing legal
restrictions (like the Fair Credit Reporting Act), which requires adverse credit information
(like a bankruptcy) to be stricken from one’s record after a certain number of years (10
in the U.S.)? Specifically, Musto p. 726 states that his empirical “results bear on the
informational efficiency of the consumer credit market, the efficacy of regulating this market
with reporting limits, and the quality of postbankruptcy credit access, which in turn bears
on the incentive to file in the first place.” He finds p. 747 “the removal of the flag leads
to excessive credit, increasing the eventual probability of default. This is concrete evidence
that the flag regulation has real economic effects. This is market efficiency in reverse.” We
use our model to assess this efficiency concern. In a world of incomplete markets and private
information, flag removal may provide insurance to impatient agents in our framework that
competitive intermediaries may not be able to provide. Hence extending the length of time
that bankruptcy flags remain on credit records may not necessarily raise ex-ante welfare. This
issue echoes Hart’s [11] examples where the opening of a market in a world of incomplete
markets may make agents worse off and Hirschleifer’s [12] finding regarding the potential
inefficiency of revealing information.
The paper is organized as follows. Section 2 describes a model economy where there
are no restrictions on information about asset market behavior, defines an equilibrium, and
discusses existence. Section 3 describes a model economy where there are restrictions on what
10Given the choice between Chapter 7 and 13, individuals would choose to file Chapter 13 only if theywished to keep assets they would lose under a Chapter 7 filing. Since borrowers in our model have negativenet worth (there is only one asset), Chapter 7 is always the preferred means to file for bankruptcy.
11This fact rules out the purchase of consumption insurance from savings in the period of discharge studiedby Bulow and Rogoff [3].
8
information on asset market behavior can be kept in an agents credit history. In particular,
we assume that information can be kept only for a finite amount of time and that there
are partitions on what asset transactions are recorded. These restrictions on information
are intended to capture the requirement that adverse events be stricken from an individual’s
credit history and the fact that credit scores are based on debt transactions rather than assets
in the current system. Section 4 estimates parameters of the model of Section 3 to match
certain key moments in the data. Section 5 studies the properties of the model. Section 6
assesses the welfare consequences of restrictions on asset market information used by credit
scoring agencies like that in the model of Section 3 compares to the unrestricted case of
Section 2. This exercise sheds some light on the impact of the Fair Credit Reporting Act.
2 Model Economy 1
2.1 People, Preferences and Endowments
Time is discrete and indexed by t = 0, 1, 2, . . . There is a unit measure of infinitely-lived
people alive at each date. At each date, a person can be one of two types, denoted it ∈ g, b.
An individual of type g (or b) at time t can become an individual of type b (or g) at the
beginning of time t+ 1 with probability Γit+1=b,it=g ∈ (0, 1) (or Γgb ∈ (0, 1)), respectively.12
Let γ denote the unconditional probability that an individual is of type H. An individual of
type it draws her endowment et independently (across time and agents) from a probability
space (Ei,B(Ei),Φi), where Ei = [ei, ei] ⊂ R++ is a strictly positive closed interval and
B(Ei) is the Borel sigma algebra generated by Ei.
Denote the life-time utility from a non-negative stream of current and future consumption
ct, ct+1, ct+2, . . . of an individual who is of type it by Ui(ct, ct+1, ct+2, . . . , θt) where θt ∈ Θ
is an independent (across time and agents) time preference shock drawn at time t from a
12This is a similar assumption to Phelan [17], who studies reputation aquisition by a government.
9
finite set with probability mass function Λ. For each i, Ui(ct, ct+1, ct+2, . . . , θt) is defined by
the recursion
Ui(ct, ct+1, ct+2, . . . , θt) = ui(ct) + βiθt∑j,θt+1
ΓjiUj(ct+1, ct+2, ct+3, . . . , θt+1)Λ(θt+1) (1)
where, for all i, ui(ct) : R+ → R is a bounded, continuous, twice differentiable and strictly
concave function and βi ∈ [0, 1).
Importantly, we assume that a person’s type it, endowment et, and time preference shock
θt are unobservable to others.
2.2 Default Option and Market Arrangement
There is a competitive credit industry that accepts deposits and makes loans to individuals.
We assume that there is a finite set L ⊂ R of possible loans or deposits (L contains negative
and positive elements as well as 0). If an individual takes out a loan `t+1 < 0 at time t
there is some probability pt that the individual will repay `t+1 units of goods at time t + 1.
If `t+1 > 0 then the individual makes a deposit which we assume that the intermediary
promises to pay back with probability pt = 1 for simplicity.
A probability of repayment pt < 1 reflects the possibility of default on the part of the
individual. We model the default option to resemble, in procedure, a Chapter 7 bankruptcy
filing. If an individual defaults, the individual’s beginning of period liabilities are set to zero
(i.e., the individual’s debt is discharged) and the individual is not permitted to enter into
new contracts in the period of default.
There is a competitive market in financial contracts. The unit price of a financial contract
(`t+1, pt) is q(`t+1, pt). For `t+1 < 0, q(`t+1, pt)·(−`t+1) is the amount received by an indvidual
at time t who promises to pay `t+1 next period with probability pt. For `t+1 > 0, q(`t+1, 1)·`t+1
is the amount handed over by the individual at time t in return for the certain promise to
10
receive `t+1 next period.
As noted earlier, there are two types of people in this economy. Let st ∈ [0, 1] be the
prior probability at time t that a person is of type g. Beliefs about an individual’s type are
important to lenders because the probability of repayment on a consumer loan may (and
will) vary across types. An important part of the market arrangement is the existence of
an agency that collects information on financial transactions of every individual and, using
this information, estimates the probability st that a given individual is of type g at time t.
We call an individual’s estimated repayment probability the individual’s credit score. We
call the agency that computes this score the credit scoring agency. And, we call the type
probability s on which the credit score (or the repayment probability) is based an individual’s
type score.13
Thus the existence of the credit scoring agency implies the presence of two functions that
are part of the market arrangement. First, there is a credit scoring function p(`t+1, ψ) which
gives the estimated probability of repayment on a loan `t+1 < 0 taken out by an individual
with type score ψ. And, second, there is a type score updating function ψ(dt, `t+1, `t, st) which
gives an individual’s type score at the start of next period conditional on having begun the
current period with asset `t and type score st and choosing (dt, `t+1) – a choice of default
corresponds to the 2-tuple (1, 0) and choice of loan/deposit `t+1 corresponds to the 2-tuple
(0, `t+1) (the precise definitions of these functions will be given in the next section).
13Nothing depends on the assumption that there are only two types. With N > 2 types, we could let stbe a N − 1 length vector (and correspondingly ψ be a vector valued function). Even in this case, the creditscore pt is just the probability of repayment on a loan.
11
2.3 Decision Problems
2.3.1 People
Let a current variable, say at, be denoted a and let next period’s variable at+1 be denoted a′.
In the special case of assets/liabilities we will let `t+1 be denoted y and `t be denoted x. Let
Y = (d, y) : (d, y) ∈ (0 × L) or (d, y) = (1, 0) be the set of possible (d, y) choices (recall
that a person can borrow or save only if she does not default and if she defaults then she
cannot borrow or save).
Each individual takes as given
• the price function q(y, p) : L−− × [0, 1] ∪ L+ × 1 → R,
• the credit scoring function p(y, s′) : L−− × [0, 1]→ [0, 1], and
• the type scoring function ψ(d, y, x, s) : Y× L× [0, 1]→ [0, 1].
We can now develop the recursive formulation of an individual’s decision problem. The
state variables for an individual are (i, e, x, s). We begin with the definition of the set of
feasible actions.
Definition 1 Given (e, x, s), the set of feasible actions is a finite set B(e, x, s; q, p, ψ) ⊂
Y that contains (i) all (0, y) y < 0 such that c = e + x − q(y, p(y, s′)) · y ≥ 0, where
s′ = ψ(d, y, x, s), (ii) all (0, y) y ≥ 0 such that c = e+ x− q(y, 1) · y ≥ 0 and (iii) if x < 0 it
also contains (1, 0).
Observe that the feasible action set does not depend on i nor θ since these are not directly
known either to financial intermediaries or to the credit scoring agency. Of course, the credit
scoring agency assigns probabilities to the individual being of a good type, ψ, and the set
of feasible actions does depend on these probabilities. The dependence of the feasible action
set on the functions p, q and ψ is noted.
12
We permit randomization so individuals choose probabilities over elements in the set of
feasible actions. We will use m(d, y) ∈ [0, 1] to denote the probability mass on the element
(d, y) ∈ Y and m as the choice probability vector.
Definition 2 Given (i, e, θ, x, s) the feasible choice set Mi(e, θ, x, s; q, p, ψ) is the set of all
m ≥ 0 such that (i) m(d, y) = 0 for all (d, y) /∈ (B(e, x, s; q, p, ψ)) and (ii)∑
(d,y)∈Ym(d, y) =
1.
In order to keep the type score updating function well defined across all actions (thereby
avoiding having to supply an exogenous set of off-the-equilibrium-path beliefs) , we will
assume each probability vector m assigns some (very small) probability ε > 0 on every
feasible action (i.e. m ≥ ε). This is similar to the “trembling hand” assumption made in
Selten [18] and can be interpreted as “tiny mistakes”.
Given (i, e, θ, x, s) and the functions p, q and ψ, the current-period return of a type i
individual from choosing a feasible action (0, y) is
Ri(0, y; e, x, s, q, p, ψ) =
ui(e+ x− q(y, p(y, ψ(0, y, x, s)) · y) if y < 0
ui(e+ x− q(y, 1) · y) if y ≥ 0
and the current-period return from choosing (1, 0) (if this choice is feasible) is
Ri(1, 0; e, x, s, q, p, ψ) = ui(e).
Denote by Wi(x, s) : L× [0, 1] → R the value of a type i individual before learning the
current realization of their endowment, time preference, and type shocks. Then, a currently
type-i individual’s recursive decision problem is given by
Vi(e, θ, x, s; q, p, ψ,Wi) = maxm∈Mi(e,x,s;q,p,ψ),m≥ε
∑(d,y)
[Ri(d, y; e, x, s, q, p, ψ)+βiθWi(y, ψ(d, y, x, s))]·m(d, y).
13
A solution exists if there exists a unique set of value functions W ∗i (·), i ∈ g, b such that
W ∗i (x, s; q, p, ψ) =
∑j∈g,b,θ
Γji
∫E
Vj(e, θ, x, s; q, p, ψ,W∗j )Φj(de)Θ(θ) ∀ i ∈ g, b.
Denote the set of optimal decision rules by M∗i (e, θ, x, s; q, p, ψ) and let mε∗
i (e, θ, x, s, q, p, ψ) ∈
M∗i (e, θ, x, s; q, p, ψ).
2.3.2 Financial Intermediary
The (representative) financial intermediary has access to an international credit market where
it can borrow or lend at the risk-free interest rate r ≥ 0. The intermediary operates in a
competitive market and takes the price function q(y, p) as given. The profit π(y, p) on
financial contract of type (y, p) is:
π(y, p) =
(1 + r)−1p · (−y)− q(y, p) · (−y) if y < 0
q(y, 1) · y − (1 + r)−1 · y if y ≥ 0(2)
Let B(L× [0, 1]) be the Borel sets of L× [0, 1]. Let A be the set of all measures defined on
the measurable space (L× [0, 1],B(L× [0, 1]). For α ∈ A, α(y, P ) is the measure of financial
contracts of type (y, P ) ∈ B(L × [0, 1]) sold by the financial intermediary. The decision
problem of the financial intermediary is:
maxα∈A
∫π(y, p) dα(y, p).
2.3.3 Credit Scoring Agency
We do not explicitly model the process by which the credit scoring agency computes type
scores and credit scores. Instead, we impose restrictions on the outcome of this process.
Specifically we assume that (i) p(y, s′) is the actual fraction of people with loan y and type
14
score s′ who repay and (ii) ψ(d, y, x, s) is the actual fraction of g type among people who
start with assets x, type score s, and choose (d, y).
Condition (i) implies
p(y, s′) = s′ ·
[∑θ′
∫[1−mg(1, 0; e′, θ′, y, s′, q, p, ψ)]Φg(de
′)Λ(θ′)
](3)
+(1− s′) ·
[∑θ′
∫[1−mb(1, 0; e′, θ′, y, s′, q, p, ψ)]Φb(de
′)Λ(θ′)
].
Condition (ii) implies
ψ(d, y, x, s; q, p, ψ) = (4)
(1− Γbg)
[Pg(d, y;x, s, q, p, ψ)s
Pg(d, y;x, s, q, p, ψ)s+ Pb(d, y;x, s, q, p, ψ)(1− s)
]+Γgb
[Pb(d, y;x, s, q, p, ψ)(1− s)
Pg(d, y;x, sp, q, ψ)s+ Pb(d, y;x, s, q, p, ψ)(1− s)
]
where
Pi(d, y;x, s, q, p, ψ) =∑θ
∫mi(d, y; e, θ, x, s, q, p, ψ)Φi(de)Λ(θ). (5)
2.4 Equilibrium
We can now give the definition of a recursive competitive equilibrium.
Definition 3. An ε-constrained recursive competitive equilibrium is a (i) a pric-
ing function q∗(y, p), (ii) a credit scoring function p∗(y, s′), (iii) a type scoring function
ψ∗(d, y, x, s; p∗, q∗, ψ∗) and (iv) a set of decision rules mε∗i (e, θ, x, s, q∗, p∗, ψ∗) such that
1. mε∗i (e, θ, x, s, q∗, p∗, ψ∗) ∈M∗
i (e, θ, x, s; q∗, p∗, ψ∗), ∀i,
2. q∗(y, p) is such that π(y, p; q∗(y, p)) = 0 ∀ y and p,
15
3. p∗(y, s′) satisfies condition (3) for mε∗i (e, θ, x, s, q∗, p∗, ψ∗), i ∈ g, b,
4. ψ∗(d, y, x, s) satisfies (4) for mε∗i (e, θ, x, s, q∗, p∗, ψ∗), i ∈ g, b.
2.5 Existence
From condition (2) in the definition of equilibrium, we know that prices q simply depend
on p and a primitive r so the equilibrium problem reduces to finding a pair of functions
p∗(y, s′) and ψ∗(d, y, x, s) such that (1) − (4) hold. To prove existence we need to establish
the following steps.
1. Since both functions share the same domain, we can extend p in the following way.
Let Ω = 0, 1 × L × L × [0, 1] and let ω be an element of Ω. For d = 0 and y < 0,
p(d, y, x, s′) = p(y, s′) for all x;for d = 0 and y ≥ 0, p(d, y, x, s′) = 1 for all x and s′;
for d = 1 and y = 0, p(d, y, x, s′) = 0 for all x and s′.
2. Stack the functions to create the continuous vector valued function
f : Ω→ [0, 1]2 ≡
f 1(ω)
f 2(ω)
≡ ψ(ω)
p(ω)
.Let F be the set of all such functions. Let ‖f‖ = maxsupω f
1, supω f2. Let K ⊂ F
be the set of functions for which (i) f 1(ω) ∈ B ⊂ [0, 1] where B is a closed interval in
[0, 1] with strictly positive elements, and (ii) f 2(ω) = 1 for all d = 0 , y ≥ 0, x and s′
and f 2(d, y, x, s′) = 0 for all d = 1, y = 0, x and s′. Then K is a closed (in the max-sup
norm), convex subset of F.
3. Define an operator T (f) : K → F in the following way. Given f ∈ K, solve the
individuals problem to get mi(d, y; e, x, s, f). Then use (4) and (3) to get T 1(f) and
T 2(f), respectively (to get T 2(f) we need to extend the the “output” function over to
Ω as in step 1 above).
16
4. Prove the following properties regarding T and K
(a) T (K) ⊂ K
(b) K is compact (i.e., if fn is a sequence in K then there exists a subsequence fnk
→ f ∈ K).
(c) T is continuous (i.e., if fn → f implies T (fn)→ T (f)).
5. Use the following version of the Schauder Fixed Point Theorem to assert the existence
of a fixed point. Let F be a normed vector space and let K ⊂ F be a non-empty,
convex, compact set. Then given any continuous mapping T : K → K there exists f
such that T (f) = f.
To establish the conditions for the Schauder Fixed Point Theorem, we use the fact that
a subset of F is compact if and only if it is bounded and forms an equicontinuous family.
Further, a sufficient condition for any subset of F to be an equicontinuous family is that each
component function of f (namely, f 1 and f 2) be Lipschitz with respect to the continuous
variable (s and s′,respectively) and the same Lipschitz constant applies to all functions in
the family.
3 Model Economy 2
Now we describe a model economy where there are restrictions on what information on
asset market behavior can be kept in an agents credit history. In particular, we assume
that information can be kept only for a finite amount of time and that there are partitions
on what asset transactions are recorded. These restrictions on information are intended to
capture the requirement that adverse events be stricken from an individual’s credit history
and the fact that credit scores are based upon data on liabilities rather assets. We also
17
assume that there are regulatory and technological reasons that restrict what credit scoring
agencies and intermediaries can observe about an individual’s priors.14
An individual’s history of asset market actions (asset choices and default decisions) at
the beginning of period t is given by (`t, hTt ) where hTt = (dt−1, `t−1, dt−2, ..., `t+1−T , dt−T ) ∈
0, 1×L×0, 1× ...×L×0, 1 ≡ HT , the set of possible histories of finite length T ≥ 1.
This definition directly incorporates the restriction that information can only be kept for a
finite number, denoted T, periods. We formalize the restrictions on observability of asset
transactions via partitions on L × HT . Because all feasible actions are taken with at least
probability ε, all feasible (`t, hTt ) are possible along the equilibrium path. Let the particular
subsets (or blocks) of the partition of L × HT be denoted ΞT = H1, ..., Hk which by the
assumptions that L and T are finite is itself a finite set. The restriction that data on assets
(which we take as `t ∈ L++) are not included in the credit scoring agency’s information set
is modelled by a measurability assumption that (`t, hTt ) is constant on each block of ΞT . As
an example, suppose that T = 1 and L = `−, 0, `1+, `2+ with `− < 0 < `1+ < `2+. Then
L×HT=1 = (0, 1), (`−, 0), (0, 0), (`1+, 0), (`2+, 0) and the measurability assumption requires
H1 = (0, 1), H2 = (`−, 0), H3 = (0, 0) and H4 = (`1+, 0), (`2+, 0). To conserve
on notation, let H(`t, hTt ) denote one of the partition blocks H1, ..., Hk. We use a similar
notation, that is A(`t+1, dt) is a partition block, to denote what an intermediary can observe
regarding an individual’s current actions (`t+1, dt). For the case where L = `−, 0, `1+, `2+,
the partition block is given (coincidentally) by A1 = (0, 1), A2 = (`−, 0), A3 = (0, 0)
and A4 = (`1+, 0), (`2+, 0).
How does this change in the environment affect decision problems? Since these informa-
tional restrictions are only on the credit scoring agency (as well as the financial intermediary
since it uses credit scores as an input into its pricing calculations), the individual’s problem
is basically identical to what we had in section 2.3.1. In particular, we simply substitute hT
for s in the individual state (i, e, θ, `, s). Note that since hT is a finite object, the state space
14For instance, prices which incorporate priors are considered proprietary and are excluded from standardcredit histories.
18
is now finite except for exogenous earnings. We can also define the endogenous measure of
individuals across the state space by µi(e, θ, `, hT ).
The informational restrictions on the credit scoring agency affect both the credit scoring
and type scoring functions in (3)-(5). In particular now the type scoring function is given by
ψ(`′, d, `, hT ) = (1− Γbg)
[Pg(`
′, d, `, hT ) · sT
Pg(`′, d, `, hT ) · sT + Pb(`′, d, `, hT ) · (1− sT )
](6)
+ Γgb
[Pb(`
′, d, `, hT ) · (1− sT )
Pg(`′, d, `, hT ) · sT + Pb(`′, d, `, hT ) · (1− sT )
]
where the prior of an agent’s type is calculated from the population distribution
sT =∑
(`,hT )∈H(`,hT )
[∫Eg
∑θ
µg(e, θ, `, hT )Φg(de)Λ(θ)
](7)
and
Pi(`′, d, `, hT ) =
∑(`′,d)∈A(`′,d),(`,hT )∈H(`,hT )
Pi(`′, d, `, hT ) (8)
Then the credit scoring function is just as before
p(`′, ψ) = ψ(`′, d, `, hT ) ·
[∑θ′
∫[1−mg(1, 0; e′, θ′, `′, hT ′, q, p, ψ)]Φg(de
′)Λ(θ′)
](9)
+(1− ψ(`′, d, `, hT )) ·
[∑θ′
∫[1−mb(1, 0; e′, θ′, `′, hT ′, q, p, ψ)]Φb(de
′)Λ(θ′)
].
As can be easily seen, the key difference from (4)-(5) simply arises from the measurability
restrictions in (7)-(8) and we use information on the distribution of agents in the economy
µ to construct the “prior” likelihood that an agent with (`, hT ) is of type g..
19
4 Moment Matching
According to the Fair Credit Reporting Act, a bankruptcy filing stays on an individual’s
credit record for 10 years. To keep the state space workable, we assume T = 2 so that a
model period corresponds to 5 years. The discount rate β for both types is based on an
annual discount rate of 0.96. The risk-free interest rate r is set to satisfy β(1 + r) = 1.
We assume the time preference shock can take two values θ ∈ 0, 1 so that agents who
receive the low shock are myopic for one period. This implies we need only pin down one
probability for each type i, namely Λi(0). The risk aversion coefficient σ is set to 2. We
assume that the “tremble” parameter is ε = 0.0001. This is the probability that agents will
play a suboptimal but feasible action by mistake. The remaining parameters are estimated,
to which we now turn.
We assume an earnings process for low and high types that is similar to that in Chatterjee,
et. al. [4]. It is assumed that each agent takes a random draw from an annual endowment
distribution conditional on her type. The distribution of annual endowments for type i is
assumed to have the following functional form:
∫ z
ei
Φ(ei|φi) = P (ei ≤ z|φi) =
[z − eiei − ei
]φi
(10)
We use simulated method of moments to choose the parameters of the endowment distribu-
tion to match the earnings gini index, mean-to-median earnings ratio, coefficient of variation
of earnings, autocorrelation of earnings, and the percentage of earnings for the second to fifth
quintiles (see the Appendix 7.3 for our estimation algorithm). We used data from the PSID
2001-2005 to construct those statistics. We restrict our sample to households whose heads
are between the age of 20 and 62. We calculate average annual earnings in the three survey
years (2001, 2003, and 2005) and then multiply that number by five to get the average five
year earnings estimate. The calculation of the autocorrelation of earnings takes advantage of
the panel properties of the dataset. It uses earnings in 2001 and 2005 of the same households.
20
The estimated support of annual endowments for type b is [0.001,0.27] (we set eb = 0.001
exogenously since in the theory since we restrict the support to be strictly positive). The
support of annual endowments for type g is [0.82,2.75]. The probability of type g switching
to type b is estimated to be 0.01, while the probability of type b switching to type g is 0.02.
This yields an invariant distribution where 2/3 of agents are type g. Table 1 summarizes the
targeted earnings statistics and the parameter values chosen to match these statistics. As
evident in the table, all parameters are statistically significant.
Table 1: Earnings Statistics (PSID 2001-2005) and Parameter Values
Statistics Target Model Parameter Estimate (s.e.)Gini index 0.45 0.45 eb 0.27 (0.00048)Mean/median 1.27 1.06 eg 0.82 (0.00140)Coeff of variation 1.33 0.80 eg 2.75 (0.00460)Autocorrelation 0.69 0.62 φg 0.50 (0.00005)Percentage of 2nd quintile 8.67 2.55 φb 0.50 (0.00311)Percentage of 3rd quintile 14.82 13.43 Γgb 0.02 (0.00008)Percentage of 4th quintile 23.46 29.95 Γbg 0.01 (0.00017)Percentage of 5th quintile 50.90 53.70
Taking the earnings parameters as given, we then estimate the remaining parameters
by matching data moments on delinquency and wealth statistics. The distribution of delin-
quency rates in Figure 1 allows us to construct the moments for overall and subprime delin-
quency rates. Other statistics, which include the debt to earnings ratio, asset to earnings
ratio, and coefficient of variation of interest rates, are obtained from the 2004 SCF.
The set of asset choices L includes one borrowing level (x), two saving levels (x1 and x2),
and zero. The borrowing level is -0.03. The two saving levels are 0.01 and 2.31. Therefore,
there are five elements in Y, which are (1, 0), (0,−0.03), (0, 0), (0, 0.01), (0, 2.31). The
probability of the time preference shock is 4.07% for type g agents and 3.86% for type
b agents. Table 2 summarizes the model statistics and the parameter values along with
standard errors. As evident in the table, all parameters except for Λb(0) are statistically
21
significant.
Table 2: Model Statistics (TransUnion and SCF) and Parameter Values
Statistics Target Model Parameter Estimates (s.e.)Targeted momentsOverall delinquency rate 29.23% 37.07% x -0.03 (0.0069)Subprime (bottom 27%) del. rate 75.74% 47.83% x1 0.01 (1.4250e-6)Debt to earnings ratio 0.002 0.002 x2 2.31 (2.5775e-6)Asset to earnings ratio 1.36 1.39 Λg(0) 0.0407 (0.0018)Coeff. of var. of interest rates 0.43 0.59 Λb(0) 0.0386 (0.0778)Untargeted momentsPercentage in debt 6.7 6.6Percentage in zero assets 7.0 5.2
5 Model Properties
Since credit scores are based on observed asset market decisions, we start by listing the
equilibrium decision rules of agents. With T = 2, there are 13 possible (x, hT ) states that
an individual will be in. If agents have the time preference shock (i.e. θ = 0), they become
perfectly myopic. In this case, they will default if they are in debt and will borrow if they
are not in debt regardless of their earnings and any other characteristics. Tables 3 and 4
provide information on equilibrium decision rules for types g and b respectively, in the event
the individual does not have a time preference shock. The tables provide the earnings levels
for a given history tuple (x, hT ) that an agent finds the action (d, y) optimal. If an action is
chosen by mistake, we list the earnings levels for which such an action is feasible (we denote
this in the table as an interval following “ε for”). If an action is neither optimal nor feasible
(i.e, agents can not choose this action even by mistake), then we place an empty set ∅ in
that cell of the table. Finally, if the decision rule does not depend on history, we group all
those (x, hT ) states into one row (which explains why there are not 13 rows in the tables.
As can be seen in Tables 3 and 4, every action is taken by some agents in equilibrium (i.e.
22
there is no state that is infeasible for both types of agents).
It is clear from Table 3, that good type agents with debt x default when they have
low earnings. They save to levels x1 and x2 the higher is their earnings. Next, good type
individuals with zero assets choose to borrow if they receive low earnings or choose to save
to levels x1 and x2 the higher is their earnings across histories hT . When they choose to
borrow, even if their asset positions are all zero, their history can affect their prices since the
intermediary can use that information to better infer their types. Therefore, their feasibility
sets are different and this affects their optimal actions given endowments. Next, good type
individuals with asset level x1 choose to dissave if they receive low earnings or choose levels
x1 and x2 the higher is their earnings across all histories. Finally, good type individuals with
asset level x2 choose to stay at level x2 no matter what their earnings.
Table 3: Equilibrium Decision rules for type g agents with θ = 1
(x, hT )\(d, y) (1, 0) (0, x) (0, 0) (0, x1) (0, x2)(x, (0, l−1, d−2)) [eg, 0.64] ε for [eg, eg] ε for [eg, eg] [0.64, 2.16] [2.16, eg]
(0, (1, x, 0) NA [eb, 0.16] ε for [eg, eg] [0.16, 2.13] [2.13, eg](0, (0, 0, 1) NA [eg, 0.16] ε for [eg, eg] [0.16, 2.13] [2.13, eg](0, (0, x, 0)) NA [eg, 0.17] ε for [eg, eg] [0.17, 2.13] [2.13, eg](0, (0, 0, 0)) NA [eg, 0.16] ε for [eg, eg] [0.16, 2.13] [2.13, eg](0, (0, xn, 0)) NA [eg, 0.16] ε for [eg, eg] [0.16, 2.13] [2.13, eg]
(x1, (0, l−1, d−2) NA [eg, 0.15] ε for [eg, eg] [0.15, 2.12] [2.12, eg](x2, (0, l−1, d−2) NA ε for [eg, eg] ε for [eg, eg] ε for [eg, eg] [eg, eg]
Moving to Table 4, bad type agents with debt x default if they have low earnings or
choose to save x1 if they have high earnings. Bad type individuals with zero assets choose to
borrow if they receive low earnings or choose no assets or to save to level x1 the higher is their
earnings across histories hT . Similar to the cases for type g agents, when agents with zero
assets choose to borrow, their prices can affect their feasibility and therefore their optimal
actions. Next, bad type individuals with asset level x1 choose to dissave if they receive low
earnings or choose level x1 the higher is their earnings across all histories. Finally, bad type
23
Table 4: Equilibrium Decision rules for type b agents with θ = 1
(x, hT )\(d, y) (1, 0) (0, x) (0, 0) (0, x1) (0, x2)(x, (0, l−1, d−2)) [eb, 0.09] ε for [0.01, eb] ε for [0.03, eb] [0.09, eb] ∅
(0, (1, x, 0) NA [eb, 0.01] [0.01, 0.03] [0.03, eb] ∅(0, (0, 0, 1) NA [eb, 0.01] [0.01, 0.03] [0.03, eb] ∅(0, (0, x, 0)) NA [eb, 0.01] [0.01, 0.03] [0.03, eb] ∅(0, (0, 0, 0)) NA [eb, 0.01] [0.01, 0.03] [0.03, eb] ∅(0, (0, xn, 0)) NA [eb, 0.01] [0.01, 0.03] [0.03, eb] ∅
(x1, (0, l−1, d−2) NA [eb, 0.01] [0.01, 0.02] [0.02, eb] ∅(x2, (0, l−1, d−2) NA ε for [eb, eb] ε for [eb, eb] ε for [eb, eb] [eb, eb]
individuals with asset level x2 choose to stay at level x2 no matter what their earnings are.
Given the earnings distributions Φi, the two decision rules imply certain properties for
the type scoring function: an observation of default is more likely to come from a bad type
individual; an observation of borrowing is more likely to come from a bad type individual
too. Since the credit scoring function depends on the type scoring function via (9), this
behavior translates into implications for credit scores. Figure 2 graphs the mapping between
type scores and credit scores. The red dotted line plots the linear regression between the
two and illustrates that there is not a perfect fit. This can be seen from equations (6)-(9).
If the equilibrium borrowing actions are independent of the state/history tuple, then there
is a direct mapping between sT and ψ in (6). Hence, the higher is sT , the higher is ψ.
Because type g agents default less often, this translates via (9) into higher p. However, since
the equilibrium borrowing action depends upon the state/history tuples (which can be seen
from the equilibrium decision rules in Table 3 and 4, the type scores do not map perfectly
into credit scores. We can nonetheless still clearly see from Figure 2 that type scores and
credit scores are highly positively correlated. The correlation coefficient weighted by the
distribution measure is 0.9327.
One way to test the model is to see if it can predict the four key properties of credit score
facts stated in section 1. To do so, the dynamics of credit scores need to be constructed.
24
Let the credit score p(x,Ψ(x, 0, `t, hTt )) ≡ ρ(x, 0, `t, h
Tt ) be the probability of repayment on a
loan of size x by an agent in state (`t, hTt ). It is possible to construct a credit score like this
even if the agent does not choose (other than by mistake) to borrow in equilibrium. Then
we can also use optimal decision rules to construct the agent’s new credit score, denoted by
ρ′(x, 0, `t+1, hTt+1), according to his optimal actions last period and his updated history tuple.
1. Interest rates fall as a person’s credit score rises.
Since higher credit simply mean a higher probability of repayment and intermediaries
earn zero profits, this implies a negative relation between credit score and interest rates
as in the data (see Figure 3).
2. Default lowers a person’s score, removal raises it.
Default lowers an individual’s type score because type b are more likely to default than
type g. Specifically, type g agents default only if they get the time preference shock in
this equilibrium, but type b agents also default if they receive low earnings. As Table
5 shows, for most histories (except for (x, 0, 0, 1)) where there is debt and subsequent
default, an individual’s credit score falls.
Table 5: Updates of credit scores after default
(`t, hTt ) µ(`t, h
Tt ) ρ(x, 0, x, hTt ) ρ′(x, 0, 0, 1, x, 0)
(x, 0, 0, 0) 0.0059 0.4350 0.4240(x, 0, 0, 1) 0.0052 0.4194 0.4240(x, 0, x, 0) ε 0.7033 0.4240
(x, 0, lt−1 > 0, 0) 0.0551 0.7063 0.4240
From Table 6, we can see that agents have higher credit scores once their default
history is erased in most cases except for when they have the state and history tuple
(lt > 0, (0, 0, 1)) and choose (0, 0) because type g agents are a lot less likely to choose
zero assets than type b agents (they either borrow when they receive a time preference
shock or simply save).
25
Figure 2: Mapping between type scores and credit scores
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75Equilibrium mapping between type scores and credit scores
type score (sT)
cred
it sc
ore
(p)
Figure 3: Credit scores and interest rates
0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.750.1
0.12
0.14
0.16
0.18
0.2
0.22
0.24Interest rates and credit scores
credit score
annu
al in
tere
st r
ate
26
Table 6: Updates of credit scores after removal of default history
ρ′(x, 0, x, 0, lt, dt−1)(`t, h
Tt ) µ(`t, h
Tt ) ρ(x, 0, lt, dt−1, 0, 1) ρ′(x, 0, 0, 0, lt, dt−1)
ρ′(x, 0, lt+1 > 0, 0, lt, dt−1)0.7033
(x, (0, 0, 1)) 0.0052 0.4406 0.51410.56770.4350
(0, (0, 0, 1)) 0.0029 0.4194 0.51410.42520.7063
(lt > 0, (0, 0, 1)) 0.0164 0.4344 0.41970.7047
As in Musto [16], we can compute the changes in percentiles of the distribution of
credit scores following a removal of the bankruptcy flag from one’s record. Musto [16]
categorized bankrupt households according to their initial post-default percentage in
the distribution of credit scores and kept track of them for ten years (the length of
time the bankruptcy record stays in their credit history by the FCRA). Since T = 2,
there is not a lot of variation in state/history tuples after default; here it is simply
(0, (1, x, 0)) and this falls within the first quintile of the distribution at 5.16%. After
two model periods when their default record is erased, an individual’s new credit score
on average increases 5% to around 13.85%. Musto found that for individuals in the first
quintile of credit scores, they jumped ahead of 5% of households post default annually.
However, these households are not the group in which people are mostly affected by
the information restriction. If we raise T > 2 we should find more heterogeneity in
post default scores which would map to Musto’s dataset better.
3. Taking on more debt (paying off debt) tends to lower (raise) credit scores.
In the model, agents may choose to be in debt because they have a time preference
shock even if they are type g agents. Therefore, there is a lot of mixing for agents who
go into debt. This makes it hard to match the prediction that increasing indebtedness
27
always lowers scores. As Table 7 shows, in some cases this prediction holds, while in
others it doesn’t. For instance, when agents dissave from positive assets to zero assets,
their credit scores drop. However, when agents actually go into debt from positive or
zero assets, their credit scores may rise. As Table 8 shows, the model predicts the
increase in credit scores after agents pay off debts for agents with state/history tuple
(x, (0, 0, 1)) but the model does poorly in other cases.
Table 7: Updates of credit scores after taking on more debt
(`t, hTt ) µ(lt, h
Tt ) ρ(x, 0, 0, hTt ) ρ′(x, 0, x, 0, 0, 0, 1)
(0, 0, 0, 0) 0.0036 0.4194 0.4350(0, 0, 0, 1) 0.0029 0.4194 0.4350
(0, 0, `t−1 > 0, 0) 5.2541e-6 0.4197 0.4350(0, 1, x, 0) 0.0245 0.4240 0.4406
(`t, hTt ) µ(lt, h
Tt ) ρ(x, 0, lt > 0, hTt ) ρ′(x, 0, x, 0, `t > 0, 0)
(`t > 0, 0, 0, 0) 0.0178 0.4252 0.7063(`t > 0, 0, 0, 1) 0.0164 0.4344 0.7063
(`t > 0, 0, `t−1 > 0, 0) 0.8065 0.7047 0.7063(`t > 0, 0, x, 0) 0.0416 0.5677 0.7063
(`t, hTt ) µ(lt, h
Tt ) ρ(x, 0, `t > 0, hTt ) ρ′(x, 0, 0, 0, `t > 0, 0)
(`t > 0, 0, 0, 0) 0.0178 0.4252 0.4197(`t > 0, 0, 0, 1) 0.0164 0.4344 0.4197
(`t > 0, 0, `t−1 > 0, 0) 0.8065 0.7047 0.4197(`t > 0, 0, x, 0) 0.0416 0.5677 0.4197
4. Scores are mean reverting.
Figure 4 graphs the average credit score given current credit scores by using the equilib-
rium decision rules 15. It can be seen that agents with lower (higher) credit scores tend to
15The average next-period credit score given (x, hT ) is calculated as∑i
[∫Ei
∑θ,(d,y) ρ(x, 0, d, y, l, d−1)mi(d, y; e, θ, x, hT )µi(e, θ, l, hT )Φi(de)Λ(θ)
]∑i
[∫Ei
∑θ µi(e, θ, l, hT )Φi(de)Λ(θ)
]
28
Table 8: Updates of credit scores after paying off debt
(`t, hTt ) µ(`t, h
Tt ) ρ(x, 0, `t, h
Tt ) ρ′(x, 0, 0, 0, x, 0)
ρ′(x, 0, lt+1 > 0, 0, x, 0)(x, 0, 0, 0) 0.0059 0.7033 0.5141
0.5677(x, 0, 0, 1) 0.0052 0.4406 0.5141
0.5677(x, 0, x, 0) ε 0.7033 0.5141
0.5677(x, 0, lt−1 > 0, 0) 0.0428 0.7063 0.5141
0.5677
Figure 4: Mean reversion of credit scores
0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
Mean reversion of credit score
credit score (p)
aver
age
next
−pe
riod
cred
it sc
ore
(p′ )
have higher (lower) credit scores next period. Therefore, the (red) linear regression line has
a flatter slope than the (blue dotted) 45 degree line.
6 Policy Experiment
Here we use the model to address a question about the welfare consequences of imposing legal
restrictions (like the Fair Credit Reporting Act), which requires adverse credit information
(like a bankruptcy) to be stricken from one’s record after a certain number of years (10 in
the U.S.). As discussed in the introduction, in a world of incomplete markets and private
29
information, flag removal may provide insurance to impatient agents in our framework that
competitive intermediaries may not be able to provide. Hence extending the length of time
that bankruptcy flags remain on credit records may not necessarily raise ex-ante welfare.
This issue is similar to Hart’s [11] examples where the opening of a market in a world of
incomplete markets may make agents worse off and Hirschleifer’s [12] finding regarding the
potential inefficiency of revealing information.
To assess this question, we compute consumption equivalents using the following formulas.
Say the PDV of utility starting in state (i, e, θ, x, hT ) for a given T is given by
v(i, e, θ, hT ;T ) = Ei
[∞∑t=0
(βθ)tct(i, e, θ, h
T ;T )1−γ
1− γ
].
To assess how much an agent in history (i, e, θ, x, hT ) would be willing to pay forever to
be in a regime where T = ∞ and there are no partitions, for each (i, e, θ, hT ) we compute
λ(i, e, θ, x, hT ) such that
v(i, e, θ, x, h∞;∞) = Ei
[∞∑t=0
(βθ)t[(1 + λ(i, e, θ, x, hT ))ct(i, e, θ, x, h
T ;T )]1−γ
1− γ
]= (1 + λ(i, e, θ, x, hT ))1−γv(i, e, θ, x, hT ;T )
or
λ(i, e, θ, x, hT ) =
[v(i, e, θ, x, h∞;∞)
v(i, e, θ, x, hT ;T )
]1/(1−γ)
− 1.
Then the total welfare gain/loss is given by
∑i,e,θ,x,hT
λ(i, e, θ, x, hT )µ(i, e, θ, x, hT ).
We use the same parameterization in the calibrated model for the T =∞ world with no
30
partitions. As a whole, the economy is better off without the legal restriction (specifically,
the welfare figure is 0.0053). This small aggregate welfare gain, however, hides the fact that
not all agents would be willing to pay to get rid of the restriction. Table ?? reports the
average consumption equivalents by types and time preference shock.
Table 9: CE by types and shocks
θ\i g b1 -0.0007 -0.00440 0.0000 -0.0013
31
References
[1] Athreya, K. (2002) “Welfare Implications of the Bankruptcy Reform Act of 1999,”
Journal of Monetary Economics, 49, p1567–95.
[2] Ausubel, L. (1999) “Adverse Selection in the Credit Card Market”, mimeo, University
of Maryland.
[3] Bulow, J. and K. Rogoff (1989) “Sovereign Debt: Is to Forgive to Forget?” American
Economic Review, 79, p. 43-50.
[4] Chatterjee, S., D. Corbae, M. Nakajima, and V. Rios-Rull (2007) “A Quantitative
Theory of Unsecured Consumer Credit with Risk of Default”, Econometrica, 75(6), p.
1525-1589.
[5] Chatterjee, S., D. Corbae, and V. Rios-Rull “A Finite-Life Private-Information Theory
of Unsecured Debt”, forthcoming, Journal of Economic Theory.
[6] Cole, H., J. Dow, and W. English (1995) “Default, Settlement, and Signalling: Lend-
ing Resumption in a Reputational Model of Sovereign Debt”, International Economic
Review, 36, p.
[7] Diamond, D. (1989) “Reputation Acquisition in Debt Markets”, Journal of Political
Economy, 97, p. 828-862.
[8] Edelberg, W. (2006) “Risk-based Pricing of Interest Rates for Consumer Loans”, Jour-
nal of Monetary Economics, 53, p. 2283-98.
[9] Fisher, J., L. Filer, and A. Lyons (2004) “Is the Bankruptcy Flag Binding? Access to
Credit Markets for Post-Bankruptcy Households”, mimeo, proceedings of the American
Law and Economics Association Annual Meetings, Berkeley Electronic Press.
[10] Gross, D. and N. Souleles (2002) “An Empirical Analysis of Personal Bankruptcy and
Delinquency”, Review of Financial Studies, 15, p. 319-47.
32
[11] Hart, O. (1975) “On the Optimality of Equilibrium when the Market Structure is In-
complete”, Journal of Economic Theory, 11, p. 418-443.
[12] Hirschleifer, J. (1971) “The Private and Social Value of Information and the Reward to
Inventive Activity”, American Economic Review, 61, p. 561-74.
[13] Kehoe, T. J., and D. Levine (1993) “Debt Constrained Asset Markets,” Review of
Economic Studies, 60, p. 865–888.
[14] Livshits, I., J. MacGee, and M. Tertilt (2007) “Consumer Bankruptcy: A Fresh Start,”
American Economic Review, 97, p. 402-18.
[15] Maroto, J. and M. Moran (2005) “Lipschitz Continuous Dynamic Programming with
Discount,” Nonlinear Analysis, 62, p. 877-94.
[16] Musto, D. (2004) “What Happens When Information Leaves a Market? Evidence From
Post-Bankruptcy Consumers”, Journal of Business, 77, p. 725-748.
[17] Phelan, C. (2006) “Public Trust and Government Betrayal”, Journal of Economic The-
ory, 130, p.27-43.
[18] Selten, R. (1975) “Reexamination of the Perfectness Concept for Equilibrium Points in
Extensive Games”, International Journal of Game Theory, 4, p. 25-55.
33
7 Appendix
7.1 Algorithm to compute T=∞ equilibrium with no partitions
1. Set grid points for endowments and scores.
(a) There are 220 endowment grid points equally spaced between the bounds of the
endowment distribution for each type.
(b) There are twenty score grid points equally spaced between ΓLH and 1− ΓHL.
2. Start iteration j = 1 with a set of initial guesses for the price function qj(y, p), the
credit scoring function pj(y, s′), and the type scoring function ψj(d, y, x, s).
3. Given the individual state (e, x, s), solve for the feasible actions set Bj(e, x, s; qj, pj, ψj).
4. Solve for V ji (e, θ, x, s; qj, pj, ψj,Wi) by value function iteration. If s′ is not on grids,
linear interpolation is used for Wi(y, s′). The solution gives the set of optimal decision
rule mji (d, y; e, θ, x, s, qj, pj, ψj) ∈M j
i (e, θ, x, s; qj, pj, ψj).
5. Given mji (e, θ, x, s; q
j, pj, ψj), calculate ψj+1(d, y;x, s, qj, pj, ψj).
6. Given ψj+1(d, y;x, s, qj, pj, ψj), calculate pj+1(y, s′) and qj+1(y, pj+1).
7. Start iteration j + 1 by using qj+1(y, p), pj+1(y, s′), and ψj+1(d, y, x, s) as the new set
of initial guesses. Repeat until they converge.
8. (optional) Solve for the stationary distribution µi(e, θ, x, s) according tomi(d, y; e, θ, x, s, q, p, ψ)
and ψ(d, y, x, s). These distribution are defined recursively by
µi′(e′, θ′′(d, y, x, s)) =
∑i,θ
(Γi′i · f(e′|i′) · Λ(θ′)
∫e
mi(d, y; e, θ, x, s, q, p, ψ)µi(Φ(de), θ, x, s)
).
(11)
34
7.2 Algorithm to compute T=2 equilibrium with partitions
1. Set grids for endowments. There are 220 endowment grid points equally spaced between
the bounds of the endowment distribution for each type.
2. Create history tuples hT=2 = (d−1, x−1, d−2). The set of history tuples is denoted as
H = (0, 0, 1), (0, x, 0), (1, x, 0), (0, 0, 0), (0, x1, 0), (0, x2, 0).
3. List all possible action/history pairs consistent with the partition that financial in-
termediaries can only observe default and borrowing. This will be useful in the later
calculations of updating functions.
We have now four tables and 43 applicable cells. A cell is marked NA if it is not
an applicable action/history pair (for instance, a household can not default with non-
negative assets which is why there are NAs in two rightmost columns in Table 1).
Because there are partitions, one cell may include more than one possible action/history
pair (for instance, in Table 1, the cell in the fifth row and first column includes the
asset/history tuples (x, 0, x1, 0), (x, 0, x2, 0)). In that case, every action/history tuple
in that same cell must have the same score due to the measurability restriction.
• Table 1: For (d, y) = (1, 0),the possible histories are
hT=2\x x 0 x1, x2
(0, 0, 1) OK NA NA
(0, x, 0) OK NA NA
(1, x, 0) NA NA NA
(0, 0, 0) OK NA NA
(0, x1, 0), (0, x2, 0) OK NA NA
• Table 2: For (d, y) = (0, x),the possible histories are
35
hT=2\x x 0 x1, x2
(0, 0, 1) OK OK OK
(0, x, 0) OK OK OK
(1, x, 0) NA OK NA
(0, 0, 0) OK OK OK
(0, x1, 0), (0, x2, 0) OK OK OK
• Table 3: For (d, y) = (0, 0), the possible histories are
hT=2\x x 0 x1, x2
(0, 0, 1) OK OK OK
(0, x, 0) OK OK OK
(1, x, 0) NA OK NA
(0, 0, 0) OK OK OK
(0, x1, 0), (0, x2, 0) OK OK OK
• Table 4: For (d, y) = (0, x1), (0, x2),the possible histories are
hT=2\x x 0 x1, x2
(0, 0, 1) OK OK OK
(0, x, 0) OK OK OK
(1, x, 0) NA OK NA
(0, 0, 0) OK OK OK
(0, x1, 0), (0, x2, 0) OK OK OK
4. Start iteration j = 1 with a set of initial guesses for the price function qj(y, p), the
credit scoring function pj(y, hT=2), and the type scoring function ψj(d, y, x, hT=2).
5. Given the individual state (e, x, s), solve for the feasible actions setBj(e, x, hT=2; qj, pj, ψj).
6. Solve for V ji (e, θ, x, hT=2; qj, pj, ψj,Wi) by value function iteration. The solution gives
the set of optimal decision rule mji (e, θ, x, h
T=2; qj, pj, ψj) ∈M ji (e, θ, x, hT=2; qj, pj, ψj).
36
7. Given M ji (e, θ, x, hT=2; qj, pj, ψj), solve for stationary distribution µji (e, θ, x, h
T=2).
µji′(e′, y, d, x, d−1) =∑
i,θ,(x−1,d−2)
(Γi′i · f(e′|i′) · Λ(θ′)
∫e
mi(d, y; e, θ, x, d−1, x−1, d−2, q, p, ψ)µi(Φ(de), θ, x, d−1, x−1, d−2)
).
(12)
8. Givenmji (e, θ, x, h
T=2; qj, pj, ψj) and µji (e, θ, x, hT=2), calculate ψj+1(d, y;x, hT=2, qj, pj, ψj)
with respect to the partition blocks listed in step 3.
9. Given ψj+1(d, y;x, hT=2, qj, pj, ψj), calculate pj+1(y, hT=2) and qj+1(y, pj+1).
10. Start iteration j + 1 by using qj+1(y, p), pj+1(y, hT=2), and ψj+1(d, y, x, hT=2) as the
new set of initial guesses. Repeat until they converge.
11. With the distribution, the type score of an agent sT (x, hT ) can be calculated according
to equation (7).
7.3 Algorithm to estimate the T = 2 model parameters
1. Let the parameter vectors to be estimated be denoted η. In the case of estimating
the parameters of the earnings process in Table 1, η = eb, eg, eg, φb, φg,Γgb,Γbg. In
the case of estimating the asset grid and time preference shock parameters in Table 2,
η = x, x1, x2,Λg(0),Λb(0). Let #(η) denote the number of parameters in η and #M
denote the number of data moments we consider.
2. Estimate η by Simulated Method of Moments (SMM). Specifically, the parameters η
are chosen to minimize
g(η) = [fD(η0)− fM(η)]W [fD(η0)− fM(η)]′, (13)
37
where fD(η0) is a vector of data moments given the true parameter vector η0, fM(η) is a
vector of model moments given parameters η, and W is any positive definite weighting
matrix.
3. We start with an identity weighting matrix and solve for a first set of parameters
η1 = arg minη g(η). These estimates are consistent but not efficient.
4. Given η1, simulate the economy with J agents for N periods, where J and N are large
numbers (J = 100, 000 and N = 1000).
(a) Starting from the stationary distribution, draw an exogenous shock processes for
type switching, time preference shock, and endowment process. Feed into the
default and asset choice decision rules to get future state. Continue for each
agent, N times to generate a sequence of exogenous shocks and decision rules.
(b) For each period, calculate #M model moments.
5. Given this #M ×N matrix of moments, calculate the associated #M ×#M variance-
covariance matrix. Call it ΣN .
6. Obtain the optimal weighting matrix as the inverse of the variance-covariance matrix
of moments, W ∗ = Σ−1N .
7. Repeat the minimization problem to solve for parameters η2.
8. The estimates η2 are asymptotically normal for fixed number of simulations.
√N(η2 − η0)→ N
(0,
(1 +
1
N
)(∂g(η2)
∂η2
Σ−1N
∂g(η2)
∂η2
′)−1), (14)
where ∂g(η2)∂η2
is approximated by numerical derivatives.
38
Recommended